From AGV to humanoid robots: development trends and solutions for robot applications

The robotics industry is at a clear turning point. In the past, robots were mostly industrial robotic arms fixed on production lines, or automated guided vehicles (AGVs) traveling along magnetic tracks or optical guides; now, with the maturity of AI, sensors, embedded computing, and connectivity technologies, robots are rapidly moving toward autonomous mobility, environmental perception, and human-robot collaboration. In particular, autonomous mobile robots (AMRs) and AGVs have become the fastest-growing subsegment of the overall robotics market, and are gradually expanding into scenarios such as warehousing, logistics, manufacturing, healthcare, agriculture, delivery, food service, and security. This article will introduce the development trends of robot applications, as well as the related solutions launched by onsemi.

From market data, different research institutions give different CAGRs due to differences in statistical scope and years, but the direction is quite consistent. The new AMR solutions material points out that from 2026 to 2031, the global related market will grow at a CAGR of about 15.3%, with the market size growing from USD 5.18 billion in 2026 to USD 10.56 billion in 2031. If broken down by model, AMRs have a CAGR of about 19% from 2023 to 2030, collaborative robots (Cobots) as high as 32%, autonomous delivery robots (ADRs) about 30.3%, and traditional industrial robots about 13.4%. Earlier data also mentioned growth rates of about 19% for warehouse robots and about 30.3% for delivery robots; AMR/AGV from 2022 to 2028 is about 16.8%, and it is believed that the 7% given by some institutions is too conservative.
 
The driving forces behind robot development include: AI and machine learning, Physical AI, advances in sensors and connectivity, improvements in embedded computing performance, increasing automation demand, labor shortages and rising labor costs, and requirements for work safety and quality. In other words, robots are no longer just "replacing manpower," but are moving toward "collaborating with people, filling labor gaps, and improving productivity and safety," which also echoes the human-robot collaboration and high-value work emphasized by Industry 5.0.
 
In terms of application scenarios, models are spreading from AGVs to AMRs, collaboration, and humanoids. Early AGVs mainly relied on magnetic tracks, optical guides, or fixed paths, with a high degree of environmental control and low perception requirements. AMRs are different: they must replan paths, detect obstacles, and avoid people and other robots in dynamic environments, so they are highly dependent on SLAM (simultaneous localization and mapping), depth sensing, vision, and edge computing. This is also why AMRs clearly use more sensors and have higher technical thresholds than AGVs.
 
At present, robot models can be roughly divided into AGVs, AMRs, hybrid robots, humanoid robots, robotic arms, and Cobots. AMR applications have expanded from warehouse transport to last-mile delivery, agriculture, inspection, healthcare, hospitality, restaurant service, security, airport transport, and even combined with robotic arms into "AMR + picking/collaboration" systems, or combined with forklifts into "AMR + forklift." Humanoid robots are regarded as the next wave of highly complex applications, combining mobile chassis, dual arms, multi-degree-of-freedom joints, and depth vision, placing higher demands on sensors, power, motor control, and real-time computing.

Despite strong growth, robot applications still face multiple challenges. The first is high upfront investment, especially for AMRs and humanoid robots, which require large amounts of sensors and computing resources; the second is employee training and process integration, as enterprises must learn to collaborate with robots; the third is functional safety and reliability, such as safety standards including ISO10218-1, obstacle avoidance redundancy, and human-robot coexistence safety; the fourth is cost, power consumption, and battery life, especially for outdoor and multi-sensor systems; the fifth is the complexity of sensor fusion, calibration, and real-time computing.
 
Looking ahead, robot applications will evolve in several directions: first, from fixed-path AGVs to autonomous AMRs, then to hybrid, collaborative, and humanoid robots; second, from indoor warehouses to outdoors, healthcare, agriculture, delivery, and public services; third, from single sensors to multimodal fusion and edge AI; fourth, from traditional buses to single-pair Ethernet and wireless connectivity; fifth, from silicon-based power devices to SiC, GaN, and higher-efficiency power and motor drives.
 
Overall, the focus of competition in the robotics industry is no longer a single component, but system-level integration capability. Whoever can provide a complete solution from depth perception, HDR vision, low-power computing, 10BASE-T1S connectivity, BLDC/GaN drives, to intelligent protection will be more likely to secure a key position in the next wave of growth in AMR/AGV, collaborative robots, and humanoid robots. This is precisely the core logic behind semiconductor solution providers actively positioning themselves in the robotics market.

Humanoid robots are at a delicate node. On the one hand, 2026 is regarded by the industry as the "first year" of mass production, with global shipments surging nearly threefold year-on-year. On the other hand, robots entering factories have so far not been widely promoted, and the core reason is that operational efficiency is inferior to human labor and generalization capability is insufficient. The gap between soaring shipments and real deployment is precisely the core proposition most worthy of attention in the current humanoid robot industry.
 
From the data, humanoid robots are in an acceleration period from the laboratory to commercialization. According to market research institution SAG's report, global humanoid robot shipments in the first half of 2026 were about 19,100 units, nearly three times the 5,100 units in the same period of 2025, and full-year shipments are expected to approach 60,000 units. The industry has now entered the "early commercialization" stage, with the focus shifting from technical demonstrations to real commercial value creation.
 
However, the rapid growth in shipments cannot hide the deep difficulties of industrialization. Before robot industrialization, the "three great mountains" that must be crossed are data, models, and applications.
 
The first mountain: data scarcity and model bottlenecks. Embodied intelligence large models have not yet reached their "ChatGPT moment." The debate between VLA (Vision-Language-Action) models and world models continues, and multiple data collection routes such as teleoperation, UMI (Universal Manipulation Interface), Ego-centric data and model, and synthetic data are in "internal strife," with no one able to persuade the others. More critically, compliant data for real physical interaction scenarios is extremely scarce, especially lacking failure data such as falls, slips, object breakage, and human intervention, which prevents models from learning exception handling; at the same time, data is not universal across different bodies, and unified annotation and alignment standards are missing.
 
The second mountain: mismatch between hardware and scenarios. Currently, about 70% of products may need a round of deep reconstruction during actual deployment to meet the real needs of scenarios. Early body development mostly followed scientific research logic, and the payload of robotic arms, end effectors, and even whether dexterous hands are needed were often not derived from the real operational needs of scenario parties. After robots enter factories, the degrees of freedom, configurations, and sensor methods of dexterous hands vary among companies, and data cannot be shared, further amplifying adaptation difficulty.
 
The third mountain: the four major thresholds of mass production supply chains. When robots move from dozens of prototypes to thousand-unit and ten-thousand-unit scales, precision, consistency, yield, cost, and capacity all surface simultaneously. The precision consistency of core joints, the lifespan and reliability of dexterous hands, the standardization of force calibration, and batch replication capability constitute the four key thresholds of mass production. Taking dexterous hands as an example, current product lifespans are generally in the hundreds of thousands of cycles, but industrial-grade continuous operation scenarios require millions or even tens of millions of cycles, "short by one to two orders of magnitude."
 
The mass production and cost issues of humanoid robots still need to be overcome. Humanoid robots require both costs close to consumer electronics and reliability close to automotive-grade, as well as motion precision far higher than traditional industrial products. This is almost an entirely new manufacturing standard. Precision motor components, force and tactile perception, and computing and control are the "high-risk" system components of humanoid robots, and actuators and perception systems face particularly prominent technical challenges.
 
The commercialization of humanoid robots is advancing in echelons along the path of "industrial manufacturing - warehouse logistics - commercial services - home scenarios." The current goal is to enable models to work reliably within specific scenario boundaries, rather than pursuing a comprehensive breakthrough in general intelligence. The "GPT moment" of embodied models is at best 3 to 5 years away; if defined as "autonomously completing new tasks in unfamiliar environments in a closed loop," it may take about 10 years.
 
The cost of humanoid robots is gradually declining, and combined with shipment volume growth, this is jointly driving the growth curve. But the real test is -- when robots walk out of showrooms and enter real factories and warehouses, can they, like humans, understand anomalies, adapt to changes, and deliver stably. This is not only a technical issue, but also a systemic battle about manufacturing capability, supply chain collaboration, and industrial patience.

SLAM is the core of AMR autonomous navigation and needs to accomplish functions such as dynamic environment modeling, map construction, obstacle detection and tracking, path planning, localization, and odometry. Traditionally, LiDAR has been the mainstream solution for SLAM, but LiDAR is costly, so in recent years there has been a clear technological shift: from LiDAR to iToF (indirect time-of-flight) and global shutter stereo vision.
 
Vision and ToF solutions each have advantages. Global shutter image sensors can avoid motion artifacts when robots move, making them suitable for high-speed navigation and barcode reading; high dynamic range (HDR) can handle mixed lighting in warehouses, factories, and outdoors, avoiding overexposure or loss of detail in dark areas. In depth sensing, dToF is suitable for long distances, while iToF is suitable for near-to-mid distances and high-resolution 3D depth. The Hyperlux ID series AF0130/AF0131 launched by onsemi integrates iToF depth sensing and some depth processing capabilities, which can reduce system power consumption and external device requirements. In addition, SWIR (short-wave infrared) can extend to 400-2100 nm wavelengths, providing new possibilities for special applications such as agriculture and semiconductor inspection.
 
Sensor fusion has also become a necessary design. A single sensor is difficult to operate stably under all lighting, weather, and scenario conditions, so systems combine LiDAR, iToF, ultrasound, radar, global shutter cameras, and IMUs, etc., to improve reliability and safety through redundancy and complementarity. This is not only the input basis for SLAM, but also the premise for functional safety and human-robot collaboration.
 
From a system perspective, an AMR can be broken down into seven major subsystems: motor drive, power distribution, sensor fusion, wired communication, LED lighting, lithium battery charging, and processor. The general block diagram includes battery charging (PFC+LLC), robot power management (battery management, DC-DC, current protection), data acquisition (imaging, ultrasound, LiDAR, force/torque sensing), processor (CPU/FPGA, memory, voltage monitoring), motor control (motor drive, SiC/GaN, MOSFET, position sensing), lighting and audio, and communication (10BASE-T1S, Bluetooth).
 
In terms of power and charging, PFC and LLC are the core of AC/DC chargers, which can improve power factor and conversion efficiency; the power tree uses Buck converters and LDOs to provide multiple voltages. The FAN6500xx series synchronous buck controllers, NCP189/NCP59801 and other LDOs, NIS3071 four-channel eFuse, and NCP81428 hot-swap smart fuse launched by onsemi constitute a complete solution from input protection to load power supply.
 
Motor drive is the main energy consumption source of robots, so BLDC control, gate drive, and power devices are crucial. onsemi provides NCD83591 three-phase gate drive, T10 series MOSFETs, GaN drive NCP51820 and iGaN family NCP58920/21/22, as well as SiC MOSFETs, which can correspond to voltages from 40V to above 650V and reduce conduction losses, switching losses, and EMI. For position sensing, there is the NCS32100 inductive position sensor, which has full accuracy at 6,000 RPM and about ±50 arc-second accuracy, and can replace some optical encoders.
 
In terms of connectivity, 10BASE-T1S is gradually replacing traditional buses such as CAN, LIN, and RS-485. It requires only a single pair of unshielded twisted-pair cable, can reach speeds of 10 Mbps, supports multi-node and PLCA collision avoidance, and can greatly reduce harness cost and complexity. onsemi's NCN26010/26000 and the new T30HMTS2500 are representatives.
 
In terms of vision, onsemi centers on the Hyperlux family: the LP series focuses on ultra-low power consumption, the LH series focuses on low light and eHDR, Hyperlux focuses on the highest dynamic range and LED flicker suppression, the SG series focuses on small-size global shutter, and the ID series focuses on iToF depth sensing. Combined with sensors such as PYTHON, XGS, AR0234CS, AR0822, ARX383, and AR0544, they can cover diverse needs from VGA to 20MP. The PRISM reference camera module ecosystem can accelerate prototype development and mass production introduction, and cooperates with platforms such as NVIDIA, Qualcomm, NXP, and AMD Xilinx.

From AGVs carrying goods along magnetic tracks, to AMRs autonomously navigating in dynamic environments, to humanoid robots attempting to enter factories and homes, the evolution of robot applications is clearly visible: this is a path from "fixed paths" to "autonomous perception," and then to "general collaboration." Supporting this path are continuous breakthroughs in perception fusion, SLAM, depth vision, high-efficiency power and motor control, single-pair Ethernet connectivity, and edge AI computing. The series of robot-related solutions launched by onsemi is a complete solution covering perception, computing, power supply, drive, and connectivity. It is shortening the distance for robots from prototype to mass production and will be an indispensable part of your human-robot collaboration in the Industry 5.0 era.